On February 21, local time, Google released Gemma, a new generation of open source model series known as "the most powerful and lightweight in the world," with an average performance far exceeding Llama 13B, winning the throne of today's open source model overlord.。

As a result, while its multi-modal large model Gemini and OpenAI are in a fierce battle in the closed source field, Google finally declares war on Meta in the open source field with Gemma.。

Sundar Pichai, CEO of Google and Alphabet, said: "Gemma has demonstrated powerful performance and from today, the model will be available worldwide and run on laptops, workstations or the Google Cloud.。"

The Gemma test results of the main rolling model are excellent.

ParametersSpecifications: Gemini homologous multi-device operation

Gemma was inspired by Gemini and developed by Google DeepMind and other teams.。

Unlike Gemini's "family bucket" route, Gemma's main "lightweight" and "high-performance" features provide 2B (2 billion) and 7B (7 billion) two parameter specifications, each scale is divided into pre-training and instruction fine-tuning two versions to meet the different needs of developers.。

The model can run on multiple mainstream device types such as laptops, desktops, IoT, mobile devices, and the cloud。The 7B version is designed for efficient deployment and development on consumer GPUs and TPUs, while the 2B version runs directly on laptops.。

Performance: the same magnitude model was"Hanging"

Gemma is built on Google's open source models and ecosystems, including Word2Vec, BERT, T5, T5X, and more.。And thanks to Gemini's shared technology and infrastructure, Gemma has 11 test scores in 18 key benchmarks such as language understanding, reasoning, and math, including MMLU and MBPP, that surpass open source models with larger parameters such as Llama 2.。

It is worth mentioning that Gemma's performance in mathematics and code ability is very outstanding, located inHugging Face Open Source Big Model RankingForefront。

Chip configuration: optimization of "self-research + foreign aid"

Gemma is said to have achieved strong general-purpose capabilities in the text field based on self-developed AI acceleration chip TPU v5e training, while possessing the most advanced understanding and reasoning skills.。Among them, 7B uses 4096 TPU v5e, 2B uses 512 TPU v5e, the data is mainly from the English data of network documents, mathematics and code.。

Not only that, Google also announced a partnership with Nvidia to use Nvidia TensorRT-LLM to accelerate Gemma's performance; Nvidia GPUs can also be used to optimize Gemma models; Nvidia's RTX chat system is also about to open up to Gemma to better protect users' data security。

Software tools: complete set of developer friendly

In addition to the model itself, Google also provides developers with a set of AI software toolkit called Responsible Generative AI Toolkit, which helps developers and researchers prioritize building safe and responsible AI applications in three areas: security classification, debugging and guidance.

• As an open model, developers and researchers worldwide can access Gemma for free through Kaggle and Colab.。
• Gemma can be reasoned and fine-tuned through multiple frameworks such as Hugging Face Transformers, allowing users to investigate Gemma's behavior and correct problems in a timely manner。
• Gemma runs on PCs and workstations, can be deployed on Google Cloud, and supports easy deployment on Vertex AI and Google Kubernetes Engine (GKE)。First-time users of Google Cloud will receive $300 in cloud credits, and researchers can claim up to $500,000 in cloud credits.。
• Terms of use provide responsible commercial use and distribution rights for all organizations, regardless of the size of the organization。

Winning the championship is not the end, Gemma will still be alone.

In summary, Gemma seems to have successfully become Gemini's "lover" in coding, data processing, architecture design, instruction optimization, reinforcement learning based on human feedback, and evaluation methods.。

However, despite its excellent performance on standard testing tasks, Google believes that further research on Gemma is needed to ensure the accuracy of its information, target alignment of the model, handling complex logical reasoning, and enhancing the model's resistance to malicious input.。

Tris Warkentin, director of product management at Google DeepMind, believes that while Gemma has outperformed its competitors in terms of six security benchmarks and human parallel assessments, it will continue to take assessments and security resolution measures commensurate with its potential risks。

The open source Gemma will undoubtedly attract a group of software engineers to develop on the basis of Google technology, enhancing the profitability and expertise of the cloud sector.。

Jeanine Banks, vice president and general manager of Developer Relations at Google Developer X, said: "It would be perfect if Google could become the sole provider of APIs and open models, offering the broadest feature set to the community.。"